CMIC Seminar: Deep Learning for Medical Image Registration - The story so far

Speaker: Yipeng Hu
UCL Contact: Dominique Drai (Visitors from outside UCL please email in advance).
Date/Time: 17 Oct 18, 13:00 - 14:00
Venue: Roberts G08

Abstract

Recent medical image registration methods based on deep neural networks have seemly converged to the so-called end-to-end learning approaches, in which the networks are trained to directly predict displacements between a given pair of images from the unprocessed image data. Besides fast inference that enables sub-second volumetric registration, some have shown superior robustness over classical methods. In particular, I will describe a new weakly-supervised formulation that we proposed to learn voxel correspondence, arguably the overarching goal of image registration. These registration networks can be trained with practical-to-obtain anatomical labels without being driven by intensity-based similarity measures, therefore they are also suitable for multimodality registration tasks. In this seminar, I will also give an overview of recent literature in image registration using deep-learning methods. I will then conclude with a discussion on the challenges and opportunities in this technical area.